Hello Tuners,

This week, we witness AI's brightest minds debate the future of regulation in California while the world of AI benchmarks sees a new twist in integrity and ambition. From fraud accusations to new, more innovative models and a French startup shaking up multimodal AI, the headlines aren't just intriguing—they're shaping the industry's future. Let’s dive into the whirlwind of developments that keep tech and policy on their toes!

• AI Godfathers Clashing For SB 1047

• Major Benchmarking Fraud: Reflection 70B 

• GPT o1 Is Here, But Where is GPT4o?

• Mistral’s Pixtral 12B, That’s A Mouthfull to Speak

AI Godfathers Clashing For SB 1047

In a showdown that could headline any AI conference, Meta's chief AI scientist, Yann LeCun and AI “godfather” Geoffrey Hinton, have found themselves on opposite sides of California’s controversial AI safety bill, SB 1047. LeCun publicly criticized the bill’s supporters, calling their view of AI’s dangers exaggerated and naive. Hinton, however, endorsed the legislation, emphasizing the risks of powerful AI models in a signed letter urging Governor Newsom to sign it into law. This clash between two 2018 Turing Award winners has thrown the deep divisions within the AI community into sharp relief, particularly around how – or even if – AI should be regulated.

SB 1047, which could soon become law, would hold developers of large-scale AI models accountable for catastrophic harm caused by failure to implement safety measures. While some see this as a much-needed safeguard in a rapidly evolving tech world, others, like LeCun, argue that the bill could stifle innovation and slow progress. With the world watching, Governor Newsom’s decision may set a precedent for AI regulation, leaving tech giants, policymakers, and public opinion teetering on the edge of their seats. Will regulation become AI's speed bump, or is it the guardrail that’ll save us from disaster?

Major Benchmarking Fraud: Reflection 70B

Matt Shumer, co-founder of OthersideAI and its writing assistant HyperWrite, has found himself in hot water after claims about his new LLM, Reflection 70B, sparked controversy. Shumer had boasted about the model’s performance, citing groundbreaking results. However, when third-party researchers couldn't replicate his results, accusations of fraud flooded social media. Shumer broke his silence with an apology on X (formerly Twitter), admitting he “got ahead of himself” but stopping short of explaining why Reflection 70B underperformed or seemed to be linked to Anthropic’s Claude 3.5 model.

The drama thickened when critics, including Nvidia’s Jim Fan, pointed out that less powerful models can ace benchmarks, suggesting Reflection 70B may not be the breakthrough it was claimed to be. Shumer’s association with Glaive AI, the synthetic data platform he used for training, further fanned the flames of scepticism. Despite his promises to investigate the discrepancies, the AI community remains unconvinced, with many questioning the transparency of his project. Whether this saga is a cautionary tale of ambition or something more serious remains to be seen.

GPT o1 Is Here, But Where is GPT4o?

OpenAI has unveiled GPT o1 with much excitement, claiming it can handle complex scientific, mathematical, and coding tasks more precisely. This new model is designed to "think harder" before generating responses, marking a significant advancement. However, it's worth noting that some features promised for GPT-4o are still pending, reminiscent of those elusive "coming soon" updates you see in the software.

The new GPT o1, while not yet equipped to browse the web or manage various file types like its predecessor, excels at solving challenging math problems. For more routine tasks, GPT-4o remains available. OpenAI has also introduced a new safety training approach for GPT o1, enhancing its ability to navigate safety rules and prevent misuse. As it starts collaborating with AI safety institutes in the U.S. and U.K., there’s hope that OpenAI will continue making meaningful progress and address those long-awaited updates.

Mistral’s Pixtral 12B, That’s A Mouthfull to Speak

Mistral AI has officially entered the multimodal AI space with the release of Pixtral 12B, a model combining language and vision processing. Pixtral 12B enables users to analyze images alongside text prompts, marking a significant step for the French AI startup, positioning itself to compete with OpenAI and Anthropic. While not publicly available via web interfaces, developers can access Pixtral 12B’s source code on Hugging Face or GitHub. Plans are to integrate the model into Mistral’s web chatbot and La Platforme API.

Though the specifics of its training data remain undisclosed, Pixtral 12B allows for analysing images of arbitrary sizes with text queries, offering flexibility in multimodal tasks. Early reports suggest the model's architecture features 40 layers, 14,336 hidden dimensions, 32 attention heads, and a vision encoder supporting 1024x1024 image resolution and 24 hidden layers for image processing. Mistral’s ongoing development signals its ambitions to challenge industry leaders, having recently raised $640 million at a $6 billion valuation and continuing to release innovative models such as Mixtral and Codestral.

LLM Of The Week

LLaMA-Omni

Researchers at the Chinese Academy of Sciences have introduced LLaMA-Omni, an AI model that supports real-time speech interaction with large language models (LLMs). Built on Meta's Llama 3.1 8B Instruct model, LLaMA-Omni processes spoken instructions and generates both text and speech responses simultaneously, with latency as low as 226 milliseconds. This breakthrough promises to revolutionize industries like customer service and healthcare, where real-time voice AI can provide more natural and efficient user experiences.

Designed to democratize voice AI, LLaMA-Omni can be trained in under three days using just four GPUs, making it accessible for startups and smaller businesses. This offers a significant opportunity for companies aiming to compete with tech giants already integrating voice capabilities into their AI assistants. While challenges like privacy concerns and the naturalness of synthesized speech remain, LLaMA-Omni represents a game-changing advancement in conversational AI, setting the stage for voice-first interfaces to become the standard in various industries.

Weekly Research Spotlight 🔍

Agentic Retrieval-Augmented Generation for Time Series Analysis

Long-context LLMs have gained traction in recent years due to their ability to incorporate significantly larger text sequences than earlier models, offering a more straightforward alternative to retrieval-augmented generation (RAG) for long-context tasks. However, contrary to the prevailing view that long-context LLMs outperform RAG in such applications, this paper highlights a critical issue: long-context LLMs can suffer from diminished focus on relevant information, leading to degraded answer quality when processing extensive sequences.

The authors introduce order-preserve retrieval-augmented generation (OP-RAG), a mechanism designed to enhance RAG’s performance in long-context question-answering tasks to address this limitation. OP-RAG improves answer quality by ensuring that the retrieved information chunks are processed to maintain their order, optimizing focus and relevance. The model displays a distinct inverted U-shaped performance curve: as the number of retrieved chunks increases, answer quality improves initially but eventually declines as too much information dilutes relevance.

Best Prompt of the Week 🎨

Today's Goal: Try new things 🧪

Acting as a Performance Planner for an Athlete

Prompt: I want you to act as a performance planner for an athlete. You will create a comprehensive daily training plan specifically designed to help the athlete prepare for the next Olympics. You will identify a target training schedule, develop key strategies and action plans, select the tools and resources for effective physical and mental conditioning, and outline any additional activities needed to ensure peak performance. My first suggestion request is: "I need help creating a daily activity plan for an athlete who is preparing for the next Olympics and aiming to maximize their performance."

This Week’s Must-Watch Gem 💎

This Week's Must Read Gem 💎